Botsi Insight

Botsi Insight transforms messy web pages into clean, summarized, and multilingual knowledge all processed locally with Gemini Nano for privacy, efficiency, and smart research.

Comment

Inspiration

We were inspired by how researchers, students, and professionals often need to gather information from multiple web pages, only to end up drowning in clutter: ads, navigation bars, and irrelevant text. We wanted to make research simpler, cleaner, and more intelligent by letting users focus only on the meaningful parts of the web.

The rise of Gemini Nano and Chrome’s built-in AI APIs gave us the perfect opportunity to show how local AI can transform this process ( without needing cloud infrastructure or compromising privacy).

What it does

Botsi Insight is a Chrome Extension that extracts and reconstructs clean, structured knowledge from any web page. It segments content into logical sections, filters out noise, and produces a clean document ready for:

  • Summarization,
  • Translation,
  • and even interactive chat with the collected knowledge.

Everything happens entirely on-device using Gemini Nano through the Prompt, Summarizer, and Translator APIs.

How we built it

We built Botsi Insight using Svelte for its lightweight reactivity and ability to inject UI elements as an iframe directly into web pages. We also use tailwind for styling, turndown.min for converting html content to markdown and chrome storage to persist information about all processing.

The workflow is as follows:

  • When a page is added to a project, its content is segmented into headings and paragraphs.
  • Each segment is sent to the Prompt API with structured output to evaluate whether it is relevant.
  • Relevant parts are merged to form a clean, readable Markdown document.

Users can then:

  • Translate the content with the Translator API,
  • Summarize it using the Summarizer API, and
  • Chat with project summaries through the Prompt API again.

All data is stored in Chrome storage, ensuring full privacy and offline usability.

Challenges we ran into

  1. Token management: Gemini Nano supports around 9,000 tokens, so we had to design an efficient segmentation strategy and avoid wasting tokens on irrelevant text.

  2. Performance: Processing a full page with multiple API calls (one per segment) required careful batching and state management.

  3. Time constraints: As with many hackathons, balancing UI polish with backend AI logic was challenging within a limited timeframe.

  4. Session management: For the chat feature, maintaining conversational state locally proved complex, so the current implementation uses session-level context built from summaries.

Accomplishments that we're proud of

  • Designing a robust cleaning algorithm that uses Gemini Nano’s structured outputs to detect important sections.
  • Building a modern and intuitive UI that integrates seamlessly into any webpage.
  • Achieving 100% client-side AI processing (no servers), no cloud dependencies.
  • Demonstrating that small, local models can deliver real value and intelligent behavior when used with smart context segmentation.

What we learned

We learned how to leverage Gemini Nano’s structured outputs for precise filtering and reasoning on-device. We also deepened our understanding of client-side AI architectures, how to optimize for token efficiency, and how to design an engaging extension that balances UX and AI logic.

What's next for Botsi Insight

We plan to expand Botsi Insight into a complete personal research assistant, capable of:

  • Building knowledge graphs across projects.
  • Generating research roadmaps and comparative analyses.
  • Supporting collaborative workspaces and export options like JSON and Markdown.
  • Integrating with larger models for deeper reasoning when online, while keeping local AI as the default.

Ultimately, we want to prove that the web doesn’t need to be noisy, it can be clean, intelligent, and personal.

Built With

Share this project:

Updates